Senior Staff Engineer (ex Director) @ MongoDB. I like building things, solving problems and playing games, occasionally at the same time. Opinions are my own.
One of the skills of a great software developer is being able to read code you’ve just written with the eyes of someone who isn't intimately familiar with it.
Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.
It’s happening faster than we thought, and the implications deserve greater attention. https://t.co/OVVPJO7VQx
AI makes good programmers 10x better: you can easily understand and contribute to more codebases and actually make an impact.
AI makes bad programmers 10x worse: you generate loads of garbage code and tests that everyone else has to understand and wade through.
some thoughts on working with ai models
• context as infra
• taste as config
• verification for autonomy
• scaling via delegation
• closing the loop
https://t.co/pdd8bk66Jz
I agree, which is why I think Mythos is effective at challenging the "impossibile" frame. The thought experiment is: "Imagine someone who sees a continuous exponential trend from vibeslop a year ago, to Opus 4.5 at year end, to Mythos and beyond. What do you think *they* might believe about the future of the SDLC in the next year or two?"
It doesn't require shifting one's own view, only being able to model someone else's. Many people won't, of course. And even if one models the other group, one doesn't have to accept the frame. But I'd like to think that an intellectually honest person would then be prompted to articulate the evidence that keeps them in the existing frame and triggers that would cause them to reconsider.
@ejames_c FWIW, I've found Mythos-class models to be a useful intuition pump. A model that finds 0-days better than humans challenges the group A frame. If the next 6-12 months brings new models that can review code better than humans, what then is the role of humans in the SDLC?
@TheZvi It's a real issue. Mine have been anti-AI indoctrinated by their school to the point where they aren't developing success skills. (Assuming non-doom futures, but 🤷)
@TheZvi Same progression from 4.5 to 4.6 of more utility but less engaging. Too quick to leap into action when it should discuss options. Subtly needs more nudges and course correction. Too literal with instructions. In creating a model that will grind for hours, they lost something.
Anthropic ARR soaring!
Also, Claude Code default effort with Opus 4.7 changed from 'medium' to 'xhigh' resulting in double token usage.
Also, 1 mil context windows rolled out, with O(n^2) cost.
Also, Enterprise contracts switched to usage-based pricing months ago.
💰💰💰
@bcherny@GergelyOrosz Today Chat would not even open a URL I provided. Search is on. I filled a support ticket. (Enterprise plan.)
I agree with Gergely. I'd be happier with fewer new features and more uptime.
@staysaasy I'm seeing multiple cases of people writing an internal CRUD app instead of sharing spreadsheets. Not quite the same thing but a lot of SaaS apps are fancy spreadsheet replacers anyway.